Autoregressive-Diffusion-Pipeline-for-Text-Generation

^{Dual pipeline for improved text generation alignment}

---

^{No link to arXiv paper yet 📖}

^{Special thanks to:}

Bergische Universität Wuppertal

_{Bergische Universität Wuppertal provided valuable learning resources and guidance on the projects direction.}

---

Abstract    Approach    Code    Results    Future Works    Credits   

Autoregressive & Diffusion LM is a hybrid framework that combines autoregressive language models (ARLMs) with diffusion-based language models (DLMs) for improved text alignment and refinement. By leveraging ARLMs for initial generation and DLMs for iterative correction, Pipeline enables more controlled, accurate, and adaptable text generation.

Abstract

Autoregressive language models (ARLMs) have achieved remarkable progress in text generation, but they still often struggle with controllability and alignment. Recent work on diffusion-based models for language modeling has introduced new possibilities for iterative refinement and text editing. In this work, we propose a hybrid framework that integrates ARLMs with diffusion-based models to improve alignment and coherence. By leveraging diffusion’s ability to iteratively refine text, our approach enables post-hoc correction and controlled editing of autoregressively generated content using auxiliary prompts. We present the architecture, training strategies, and experimental results demonstrating improved alignment and adaptability in text generation.

Approach

Our proposed Pipeline framework integrates autoregressive language models (ARLMs) with diffusion-based language models (DLMs) to enhance text generation through iterative refinement. This hybrid approach leverages the strengths of both paradigms: ARLMs excel in fluent, high-quality text generation, while DLMs offer a powerful mechanism for controlled editing and correction.

At a high level, Pipeline operates in two stages:

• Autoregressive Generation: An ARLM produces an initial text output based on a given prompt. This stage ensures efficiency and fluency in generating coherent sequences.
• Diffusion-based Refinement: A diffusion language model takes the ARLM-generated text and iteratively refines it using a secondary conditioning prompt. This allows for correction, alignment, and enhanced controllability while preserving fluency. To achieve seamless integration, we explore latent-space diffusion, where text representations are iteratively denoised rather than modifying raw tokens directly. This enables smooth, targeted refinements while maintaining contextual consistency. Additionally, we introduce alignment-aware conditioning, where the diffusion model receives both the original prompt and a secondary guidance prompt to enforce desired modifications without excessive drift from the ARLM’s intent.

Our approach balances generation speed, coherence, and flexibility, making it well-suited for applications requiring controllable text generation, post-generation alignment, and iterative text editing. We demonstrate the effectiveness of Pipeline across tasks such as content refinement, bias mitigation, and controlled rewriting, hopefully showing significant improvements over standalone ARLMs.

Code

To Be Published

Results

To Be Determined

Future works

Credits